llms.txt is a file almost nothing reads yet. That is a strange thing to recommend, and there is still a reasonable argument for it.
What it actually is
The proposal, from Jeremy Howard in 2024, is a file at /llms.txt written in Markdown. Not a directive file like robots.txt — a summary. It describes what the site is, then lists the pages that matter with a line of context each.
The reasoning is that an HTML page is mostly navigation, scripts and boilerplate, and a model given a curated Markdown index of a site can find and cite the right page more reliably than one crawling the rendered pages.
It is a sensible idea. It is also, at the time of writing, a convention rather than a standard, with no formal adoption from OpenAI, Anthropic, Google or Perplexity.
The honest state of adoption
Being straight about this matters, because a lot of SEO writing implies more than is true.
| Claim | Status |
|---|---|
| An official web standard | No — a community proposal |
| Read by ChatGPT, Claude or Gemini | Not confirmed by any of them |
| Affects Google rankings | No |
| Affects AI Overviews | No evidence |
| Costs anything to add | Minutes, if generated |
| Can hurt you | Only if it contradicts your real content |
So the case is not "this will get you cited". It is "this is nearly free, it is well-formed, and if adoption arrives you already have one". That is a weaker argument than most articles make, and it is the true one.
The part that does help today
Writing an llms.txt forces a useful exercise even if nothing ever reads the file.
To produce one you have to state, in a paragraph, what your site is for. Then you have to pick which pages matter and describe each in a line. Most sites cannot do this without discovering that their own navigation does not reflect it, or that three pages cover the same topic, or that the page they consider most important is four clicks deep.
That is a genuine content audit disguised as a config file. The output happens to be machine-readable.
What to put in it
The convention is loose. A workable shape:
- An H1 with the site name, then a blockquote summarising what the site does in two or three sentences — including anything unusual about it.
- Sections by category, each a list of links in the form
- [Title](url): one line of description. - A section for editorial if you have one, so a model answering a how-to question can find the guide rather than only the tool.
- A short policy line stating how you would like content cited.
Keep it generated rather than hand-written. A file that describes 131 pages will be wrong within a month if a human maintains it — this site emits its llms.txt from the same data that builds the pages, so it cannot drift.
What actually gets you cited
If the goal is being quoted by an AI assistant rather than merely ranked, the things that demonstrably matter are unglamorous and none of them is a file at the root.
- Answer the question in the first paragraph. Extraction favours a direct answer near the top over one buried in section four.
- Use question-shaped headings. They map onto how people ask and how models chunk.
- Be specific and checkable. Numbers, thresholds and named standards get quoted; adjectives do not.
- Structured data. FAQPage, HowTo and Article schema are read today, by systems that exist now.
- Let the crawlers in. Blocking GPTBot or ClaudeBot in robots.txt while hoping to be cited is a contradiction worth checking for.
That last one catches people out. The Robots.txt and llms.txt Generator writes both files together, so you can see what you are allowing and what you are describing in one place.
Frequently asked questions
Is llms.txt an official standard?
No. It is a community proposal from 2024 that has not been adopted by any standards body, and no major AI provider has confirmed they read it. Articles implying otherwise are ahead of the evidence.
Will llms.txt improve my Google rankings?
No. It has no role in Google's ranking systems and no demonstrated effect on AI Overviews. If someone is selling llms.txt as a ranking factor, that is not supported by anything public.
Should I add one anyway?
If it can be generated as part of your build, yes — it costs minutes and you are ready if adoption arrives. If it means hand-maintaining a file that will silently go stale, the case is much weaker.
How is it different from robots.txt or a sitemap?
robots.txt says what crawlers may access; a sitemap lists every URL for discovery. llms.txt is neither — it is a curated, human-readable summary of what matters and why, written in Markdown rather than XML.
What actually helps with AI citation?
Answering the question directly in the first paragraph, question-shaped headings, specific checkable facts rather than adjectives, and FAQPage or HowTo structured data — which is read today by systems that exist now. And not blocking the crawlers you hope will cite you.
Can llms.txt hurt my site?
Only if it contradicts your actual content or lists pages that no longer exist. A generated file cannot drift; a hand-written one describing a hundred pages will be wrong within a month.